铁道科学与工程学报2025，Vol.22Issue(9)：4218-4233,16.DOI:10.19713/j.cnki.43-1423/u.T20241979
双域多尺度特征提取的轨道面瑕疵检测算法
Dual-domain multiscale feature extraction algorithm for track surface defect detection
摘要
Abstract
Track surface defect detection is a critical technology for ensuring the safe operation of railway systems.To address the problems of low accuracy and high omission rates in existing track surface defect detection algorithms,a dual-domain multiscale feature extraction algorithm for track surface defect detection was proposed based on the YOLOv5s framework.Firstly,a dynamic enhanced upsample module was designed to reduce resolution loss and artifacts during the upsampling,thereby improving the ability to capture fine-grained features of track surface defects.Secondly,a self-synergistic convolution block attention model was proposed,which combined the advantages of self-attention and convolution block attention mechanisms to capture global context information of track surface defects while suppressing interference from irrelevant background noise.Thirdly,the omni-dimensional dynamic convolution was used to replace the standard convolution in the backbone network,achieving multi-scale feature extraction for track surface defects by dynamically adjusting the kernel parameters.Finally,a wavelet transform pyramid module was constructed to jointly extract spatial and frequency domain features of defects through Haar wavelet decomposition,enhancing global shape modeling and detail representation.Experimental results demonstrate that each improvement effectively enhances the detection performance of the model.On the self-built track surface defect dataset,the proposed algorithm achieves mAP50 and mAP50-95 of 84.4%and 53.3%,with GFLOPs of 13.8G.Compared with YOLOv5s,the mAP50 and mAP50-95 are improved by 4.6 percentage points and 6.7 percentage points,while GFLOPs decrease by 13.8%.The proposed algorithm outperforms mainstream detection models(e.g.,Faster R-CNN,RT-DETR,SSD,YOLOv7)and other track surface defect detection methods in accuracy.It also demonstrates strong generalization ability on public track surface defect datasets,proving its effectiveness in the field of track surface defect detection.关键词
轨道面瑕疵检测/上采样模块/注意力机制/全维动态卷积/小波变换Key words
track surface defect detection/upsample module/attention mechanism/omni-dimensional dynamic convolution/wavelet transform分类
信息技术与安全科学引用本文复制引用
胡贺南,都业辉,李荣华,王大志,张然..双域多尺度特征提取的轨道面瑕疵检测算法[J].铁道科学与工程学报,2025,22(9):4218-4233,16.基金项目
辽宁省教育厅科学研究项目重点项目(LJKZ0475) （LJKZ0475）
辽宁省教育厅科学研究项目(LJ212410150036) （LJ212410150036）
大连市高层次人才创新支持计划(2022RJ03) （2022RJ03）
